Introduction. Cardiovascular diseases are a leading cause of mortality globally, and advanced diagnostic tools are needed. Electrocardiographic imaging (ECGI) has emerged as a promising non-invasive technique for reconstructing epicardial potentials, providing valuable insights into cardiac conditions such as premature ventricular complexes (PVCs), which affect a significant portion of the population. The accurate localization of PVCs remains challenging due to signal attenuation and the ill-posed nature of the cardiac inverse problem. Materials and Methods. We propose a spatiotemporal Mercer kernel framework to improve PVC origin estimation from ECGI data. The model leverages spatial and temporal correlations to address ECGI challenges. Performance is compared with traditional methods, including Tikhonov regularization, using real data from the EDGAR Time Signal Catalog. An ADAM optimizer processes multiple signal frames, while cross-validation selects optimal spatial and temporal kernel parameters. Experiments and Results. The Gaussian spatiotemporal kernel outperforms both the Laplacian kernel and the Tikhonov method, achieving the lowest NSME of 0.3439 compared to 0.3503 and 0.5036, respectively. In terms of localization error relative to the pacing site, the Gaussian kernel achieves the best performance (12.55 mm), followed by the Laplacian (23.99 mm), while Tikhonov yields the largest error (36.48 mm). Conclusions. The Gaussian kernel-based approach offers a favorable direction for enhancing ECGI accuracy, improving PVC localization, and offering potential benefits for guiding catheter ablation procedures in clinical settings.